Upload
iwan-priambodo
View
216
Download
0
Embed Size (px)
Citation preview
7/27/2019 Learning Motivation.pdf
1/8
Self-Assessment of Motivation: Explicit and Implicit
Indicators in L2 Vocabulary Learning
Kevin Dela Rosa, Maxine Eskenazi
Language Technologies Institute, Carnegie Mellon University,
5000 Forbes Avenue, Pittsburgh, Pennsylvannia, USA
{kdelaros,max}@cs.cmu.edu
Abstract. Self-assessment motivation questionnaires have been used in
classrooms yet many researchers find only a weak correlation between answers
to these questions and learning. In this paper we postulate that more direct
questions may measure motivation better, and they may also be better
correlated with learning. In an eight week study with ESL students learning
vocabulary in the REAP reading tutor, we administered two types of self-
assessment questions and recorded indirect measures of motivation to see which
factors correlated well with learning. Our results showed that some user actions,
such as dictionary look up frequency and number of times a word is listened to,
correlate well with self-assesment motivation questions as well as with how
well a student performs on the task. We also found that using more direct self-
assesment questions, as opposed to general ones, was more effective in
predicting how well a student is learning.
Key words: Motivation Modelling, Intelligent Tutoring Systems, Computer
Assisted Language Learning, Motivation Diagnosis, English as a Second
Language
1 IntroductionMotivation modelling and its relation to user behavior has receieved attention by the
educational computing community in recent years. William and Burden define
motivation as a state of cognitive and emotional arousal which leads to a conscious
decision to act, and which gives rise to a period of sustained intellectual and/or
physical effort in order to attain a previously set goal (or goals)
[1]. The use of self-assessment questionaires is one common approach to measuring motivation. One such
construct is Motivated Strategies for Learning Questionnaire (MSLQ), an 81-item
survey designed to measure college students' motivational orientations and their use
of various learning strategies [2]. While questionnaires are useful to detect enduring
motivational traits, some are criticized, particularly those administered prior to
interaction. Since a students motivation is likely to change during an interaction, it is
important to use them with other methods to adapt instruction and to gather more
transient information about a students motivation [3]. Other methods of assessing
7/27/2019 Learning Motivation.pdf
2/8
2 Kevin Dela Rosa, Maxine Eskenazi
motivation include direct communication with students, emotion detection, and
recorded interactions with an intelligent tutor. For modelling and understanding user
behavior automatically, Baker [4] showed that machine learning models trained on
log data of student activity can be used to automatically detect when a student is off-
task. A study by Cetintas et al. [5] reached a similar conclusion, using a regression
model personalized to each student. And, Baker et al. [6] showed that a latent
response model can be used to determine ifa student is gaming the system in a way
that leads to poor learning.
An important issue has been how to automatically detect a students current
motivational state. As mentioned above, one method of measuring motivation is
questionnaires that cover a variety of motivation aspects. One important consideration
is how detailed and/or direct these survey questions should be with respect to the task
or, in other words, is it better to have questions that are tightly focused on the tasksbeing performed by the student or is it better to construct questions that are more
general and can cover many difference aspects of motivation, such as the MLSQ.
Also, a students usage of a tutoring system, as indicated by the amount of activity
and types of actions taken, may furnish good implicit indicators of a students
motivation.
We propose that in a computer-assisted L2 language learning environment certain
recorded student interactions during learning activities can act as implicit indicators of
that students motivation. We also propose that these implicit indicators, as well as
explicit ones like self-assessment surveys, can be used to predict the amout of
learning that is taking place. Lastly we postulate that more direct questions may
measure motivation better and may also be better correlated with learning.
For this study we used a web-based language tutor called REAP [7]. REAP, whichstands for REAder-specific Practice, is a reading and vocabulary tutor targeted at
ESL students, developed at Carnegie Mellon University, which uses documents
harvested from the internet for vocabulary learning. REAPs interface has several
features that help to enhance student learning. One key feature in REAP is that it
provides users with the ability to listen to the spoken version of any word that appears
in a reading, making use of Cepstral Text-to-Speech1 to synthesize words on demand
when they are clicked on. Additionally, students look up the definition of any of the
words, during readings, using a built-in electronic dictionary. REAP also
automatically highlights focus words, the words targeted for vocabulary acquisition in
a particular reading. REAP is a language tutor and a testing platform for cognitive
science studies [8, 9], as is the case of this study.
In this paper we describe a classroom study that compares the effectiveness of
different motivational indicators in a vocabulary learning environment. We define the
different types of survey questions we used as explicit measures of motivation and the
various user actions we recorded as indirect indicators of motivation. Next we
describe the results of a classroom study that integrated our various motivation
indicators and how well they correlated with our learning measures. Finally we
discuss the implications of our results and suggest future directions.
1 Cepstral Text-to-Speech. http://www.cepstral.com
7/27/2019 Learning Motivation.pdf
3/8
Self-Assessment of Motivation: Explicit and Implicit Indicators 3
2 Classroom StudyIn order to determine which of our hypothesized indicators of motivation were most
related to learning we conducted a classroom study with a web-based tutor, focused
on L2 English vocabulary learning, and recorded responses to motivation
questionnaires and user actions that we log, which may indirectly indicate a students
motivation level. The classroom study consisted of a pre-test and post-test with
multiple choice fill-in-the-blank vocabulary questions, and six weekly readings, each
followed by practice vocabulary questions similar to, but not the same as those in the
pre-test and post-test. During the pre-test and post-test, a set of seventeen self-
assessment motivation questions were administered, and after each weekly session
there were a set of five motivation questions.
21 intermediate-level ESL college students at the University of PittsburghsEnglish Language Institute participated in the study and completed all of the
activities. For this study the readings and vocabulary questions had 18 focus words,
taken from either the General Service List 2 or the Academic Word List3, and not part
of the class core vocabulary list.
In the following subsections we describe the types of questionnaires administered,
the recorded user actions, and the metrics we used to measure student learning.
2.1 Motivation QuestionnaireWe administered motivation survey questions as explicit measures of motivation after
each reading. 17 survey questions were administered in the pre-test and post-test, as
shown in Table 1, using a five-point Likertscale, with a response of 5 indicating thegreatest agreement with the statement and 1 indicating the least agreement. The 17
questions were divided into two groups: GeneralandDirect.
We call General survey questions high-level survey questions which have been
used in past REAP studies because of their generality; they are used in many studies
in the Pittsburgh Science of Learning Center4. For example, one of the General
questions we used was, When work was hard I either gave up or studied only the
easy parts, which can be used for many different subject matters. The design of these
questions was guided by the MLSQ [2], and aimed to use the fewest number of
questions possible that cover the most motivational constructs.
We call Directquestions the more explicit items that focused on aspects directly
related to the reading activities accomplished over the course of the study. An
example of a Direct question: Learning vocabulary in real documents is aworthwhile activity. This is focused on the specific REAP tasks.
A total of twelve General and five Direct self-assessment motivation questions
were administered during the pre-test and post-tests. Additionally, the five Direct
motivation survey questions in Table 2 were asked after each weekly reading activity,
2 The General Service List. http://jbauman.com/gsl.html3 The Academic Word List. http://www.victoria.ac.nz/lals/resources/academicwordlist4 Pittsburgh Science of Learning Center (PSLC). http://www.learnlab.org
7/27/2019 Learning Motivation.pdf
4/8
4 Kevin Dela Rosa, Maxine Eskenazi
at regular intervals in between the pre-test and post-test, to see how the responses
correlated with student behavior and learning at each reading. We wanted to
determine if there was a difference in how well each of these two question groups
correlates to the learning measures we recorded (multiple choice questions). We
hypothesize that questions more directly related to the tasks/activities performed will
be better at predicting motivation and learning. This is guided by unpublished results
of past REAP studies which have shown that higher-level questions generally failed
to correlate well with learning measures, and previous success with direct questions
by Heilman et al. [10].
Table 1. Pre-test/Post-test Motivation Survey Questions
ID Survey Question Prompt Group Type
S1 I am sure I understood the ideas in the computer lab sessions. General E
S2 I am sure I did an excellent job on the tasks assigned for the computer labsessions.
General E
S3 I prefer work that is challenging so I can learn new things. General A
S4 I think I will be able to use what I learned in the computer lab sessions in my
other classes.
General V
S5 I think that what I learned in the computer lab sessions is useful for me to know. General V
S6 I asked myself questions to make sure I knew the material I had been studying. General O
S7 When work was hard I either gave up or studied only the easy parts. General A
S8
I find that when the teacher was talking I thought of other things and didn't reallylisten to what was being said. General A
S9
When I was reading a passage, I stopped once in a while and went over what I
had read so far. General O
S10 I checked that my answers made sense before I said I was done. General O
S11 I did the computer lab activities carefully. General E
S12 I found the computer lab activities difficult. General A
S13 I continued working on the computer lab activities outside the sessions. Direct A
S14 I did put a lot of effort into computer lab activities. Direct A
S15 I did well on the computer lab activities. Direct E
S16 I preferred readings where I could listen to the words in the document. Direct V
S17 Learning vocabulary in real documents is a worthwhile activity. Direct V
Table 2. Post-reading Survey Motivation Questions
ID Survey Question Prompt Type
Q1 Did you find the spoken versions of the word helpful while reading this document? V
Q2 Do you find it easy to learn words when you read them in documents? E
Q3 Did you find this document interesting? V
Q4 Did you learn something from this document? V
Q5 Does reading this document make you want to read more documents? A
Furthermore, for this study we grouped the questions into three types:
Affective (A): Deal with emotional reactions to a task Expectancy (E): Deal with beliefs about a students ability to perform a task Value (V): Deal with goals and beliefs about the importance and interestof a
task
The 3 groups are based on the Pintrich and De Groot components of motivation
(self-efficacy, intrinsic value, test anxiety) [11]. We used this grouping to simplify the
analysis of the results. Note that in the tables and figures, Other (O) signifies a
7/27/2019 Learning Motivation.pdf
5/8
Self-Assessment of Motivation: Explicit and Implicit Indicators 5
question that failed to group into one of the three types, typically a question on
learning strategies.
2.2 Recorded User InteractionsIn addition to survey questions, we recorded actions taken by the students which we
hypothesize would indirectly correspond to motivation and might also correlate with
learning. The following were recorded during each activity:
Word lookup activity, using our built-in electronic dictionary
A1: Total number of dictionary lookups
A2: Number of focus words looked up in the dictionary
A3: Number of dictionary lookups involving focus wordsWords listening activity, using our built-in speech synthesis
A4: Mean number of listens per word
A5: Total number of listens
A6: Number of words listened to
Average time spent on activity tasks
A7: Time spent reading the documents
A8: Time spent on practice questions
2.3 Learning MeasuresIn order to assess how well students learned the target vocabulary words, we recorded
the following measures:
L1: Average post-reading practice question accuracy (for all questions appearing
directly after reading the documents)
L2: Pre-test to post-test normalized gain
L3: Post-test accuracy
L4: Average difference between pre-test and post-test scores
Note that that L2 and L4 are two different ways of looking at the improvements made
by students over the course of the study, where L2 is tuned to the relative difference
between the test scores, and L4 is sensitive to the absolute difference in scores.
3 ResultsThe results of our study show that the use of the REAP system significantly helpedstudents improve their performance on the vocabulary tests, as evident in the average
overall gains between the pre-test and post-test {L3} (p < 0.004), whose average
scores were 0.3439 ( 0.0365) and 0.5000 ( 0.0426) respectively. The average post-
reading practice question accuracy was 0.8417 {L1} ( 0.0466). The overall average
normalized gain {L2} between pre-test and post-test was 0.2564 ( 0.0466), and
average difference in score {L4} between the pre-test and post-test was 0.1561 (
0.0232).
7/27/2019 Learning Motivation.pdf
6/8
6 Kevin Dela Rosa, Maxine Eskenazi
Fig. 1. Significant correlations values between motivational & learning factors, and betweenimplicit & explicit motivation indicators. Color signifies level of significance, with green
representing strong statistical significance (p < 0.05), and yellow representing moderate
signifance (p < 0.1). Note that self-correlations and correlations with low significance values
were omitted. Also note that Q1-Q5 correspond to the students' average survey response values
for all reading activities, and with respect to the implict indicators, A1-A8 correspond to the
students' average values of those indicator values over all reading activties.
In order to find the motivational factors that best correlated with learning, we
computed a Pearson correlation coefficient matrix for the values for the different
factors and grouped question responses, and determined the significance of each pair
of correlations using a two-tailed test. Figure 1 summarizes correlation and
signifcance values found between the motivational and learning factors, and shows
how the indirect indicators correlate with the expicit indicators of motivation.Additionally, when we looked at the post-reading accuracies of each individual
reading activity (as oppose to the overall averages, which were shown in Figure 1),
the following motivational factors tended to significantly correlate with the post-
reading practice question accuracies at significance levels varying levels over the six
readings betweenp < 0.01 andp < 0.05:
Q2 response: Do you find it easy to learn words when you read them indocuments?
7/27/2019 Learning Motivation.pdf
7/8
Self-Assessment of Motivation: Explicit and Implicit Indicators 7
Q3 response: Did you find this document interesting? Total number of dictionary lookupsNumber of focus words looked up in the dictionary Time spent on practice questions
4 DiscussionWe see in Figure 1 that most of the Generalquestions did not significantly correlate
with the various learning measures, while the Directquestions did. In fact, the only
sub-group ofGeneralquestions that correlated with the learning measures was Other,
those questions that did not fit well into our three types (Affective, Expectancy, and
Value), which is not surprising since the Otherquestions mainly focused on learningstrategies as opposed to motivation. Furthermore, all of the direct questions asked
after each reading, except forQ5 (Does reading this document make you want to read
more documents?), had significant correlations with the learning measures. Perhaps
the reason Q5 failed to have significant correlations is due to the fact students were
not given the option to actually act on the desire to read more documents in our tutor
during the semester, due to class constraints. Therefore, our results imply that General
questions are less effective at predicting the learning outcome of a student than are the
Directmotivation questions which are more closely focused on the tasks performed.
Moreover, our results hew closely to past results by Bandura [12] in the domain of
self-efficacy.
Additionally, the implicit motivation indicators, based on recorded student actions,
seemed to correlate well with our learning measures, particularly our word listening
and dictionary lookup-related interactions, which implies that these kinds of actions
can also help in predicting a students motivational state. Interestingly, the amount of
time spent on reading and answering questions did not correlate well with the learning
factors, which implies that simply using the absolute amount of time spent on task
may not be a good factor to use in predicting a students learning outcomes, and
perhaps taking into account how the time was used by students would be a better
factor to consider. Furthermore, most of the implicit indicators we recorded had
significant correlations with one or more of the direct motivation survey questions in
Figure 1 that we asked after each reading, which implies that implicit indicators may
be effective in predicting the motivational state of the student using an intelligent
tutor on an activity-by-activity basis.
5 ConclusionUnderstanding and modeling motivation and student behavior is an important issue
for intelligent tutoring systems. We proposed that some student interactions recorded
by tutors can act as implicit indicators of motivation, and that these implicit
indicators, as well as explicit ones like self-assessment motivation questions, can be
used to predict student learning. We tested our hypothesis with a classroom study
using a vocabular tutor that integrates implicit and explicit indicators. The results
7/27/2019 Learning Motivation.pdf
8/8
8 Kevin Dela Rosa, Maxine Eskenazi
show that some user actions, such as dictionary look ups and listening to words,
correlate well with motivation questions and student performance. We also found that
the use ofDirectquestions, specifically tailored to the tasks, rather than Generaland
all-encompassing questions, was more effective in predicting student performance.
Acknowledgements. This project is supported by the Cognitive Factors Thrust of the
Pittsburgh Science of Learning Center which is funded by the US National Science
Foundation under grant number SBE-0836012. Any opinions, findings, and
conclusions or recommendations expressed in this material are those of the authors
and do not necessarily reflect the views of the NSF. The authors would like to thankBetsy Davis, Tamar Bernfeld, and Chris Ortiz at the University Pittsburghs ELI for
their participation and input in the studies.
References
1. Williams, M., Burden, R. L.: Psychology for language teachers: A social constructivistapproach. Cambridge University Press, Cambridge,(1997)
2. Pintrich, P.R., Smith, D.A.R., Garcia, T., McKeachie, W.: A manual for the use of themotivated strategies for learning questionnaire (MSLQ). Report, Ann Arbor (1991)
3. de Vincente, A., Pain, H.: Motivation Diagnosis in Intelligent Tutoring Systems. In: 4thInternational Conference on Intelligent Tutoring Systems (1998)
4. Baker, R.S.: Modeling and understanding students' off-task behavior in intelligent tutoringsystems. In: 25th ACM SIGCHI Conference on Human Factors in Computing Systems,
pp. 10591068. ACM, New York (2007)
5. Cetintas, S., Si, L., Xin, Y.P., Hord, C.: Automatic Detection of Off-Task Behaviors inIntelligent Tutoring Systems with Machine Learning Techniques. IEE Trans. Learn. Tech.3, 228-236 (2010)
6. Baker, R.S., Corbett, A.T., Koedinger, K.R.: Detecting Student Misuse of IntelligentTutoring Systems. In: Lester, J.C., Vicari, R.M., Paraguau, F. (eds.) ITS 2004. LNCS,
vol. 3220, pp. 5476. Springer, Heidelberg (2004)
7. Heilman, M., Collins-Thompson, K., Callan, J., Eskenazi M.: Classroom success of anIntelligent Tutoring System for lexical practice and reading comprehension. In: 9thInternational Conference on Spoken Language (2006)
8. Dela Rosa, K., Parent, G., Eskenazi, M.: Multimodal learning of words: A study on the useof speech synthesis to reinforce written text in L2 language learning. In: 2010 ISCA
Workshop on Speech and Language Technology in Education (2010)
9. Kulkarni, A., Heilman, M., Eskenazi, M., and Callan, J.: Word Sense Disambiguation forVocabulary Learning. In: 9th International Conference on Intelligent Tutoring Systems
(2008)
10. Heilman, M., Juffs, A., Eskenazi, M.: Choosing Reading Passages for VocabularyLearning by Topic to Increase Intrinsic Motivation. In: 13th International Conference on
Artificial Intelligence in Education (2007)
11. Pintrich, P.R., De Groot, E.V.: Motivational and Self-Regulated Learning Components ofClassroom Academic Performance.J. of Ed. Psych.82, 3340(1990)
12. Bandura, A.: Self-Efficacy. In: Ramachaudran, V.S. (ed.) Encyclopedia of HumanBehavior, vol 4, pp. 7181. Academic Press, New York (1994)